def thread_scaling_test(target_dir,
exp_name,
group='',
params={},
dry_run=0,
is_tgs_only=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
is_dp = 1
cs = 4096
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if (machine_info['hostname'] == 'Haswell_18core'):
k_perf_order = {0: 1500, 1: 5000, 4: 400, 5: 2000, 6: 100}
elif (machine_info['hostname'] == 'IVB_10core'):
k_perf_order = {0: 1200, 1: 3000, 4: 350, 5: 1500, 6: 80}
k_time_scale = {
n: desired_time * k_perf_order[n]
for n in k_perf_order.keys()
}
points = dict()
points[0] = [896]
points[1] = [896]
points[4] = [768]
points[5] = [768]
points[6] = [384]
count = 0
th_l = [machine_info['n_cores']]
th_max = machine_info['n_cores']
for ts, tgs_r in [(0, 0), (2, 1), (2, -1)]:
for kernel, mwdt in [(0, 1), (1, 2), (4, 1), (5, 2)]: #, (6,1)]:
for N in points[kernel]:
tb, nwf, tgs, thx, thy, thz = (-1, -1, tgs_r, -1, -1, -1)
# key = (mwdt, kernel, N, tgs_r, group)
# if key in params.keys():
# continue #already computed
if tgs == 1: mwdt = -1
key = (mwdt, kernel, N, tgs_r, 'MEM')
if key in params.keys(): #reuse the conf of existing test
tb, nwf, tgs, thx, thy, thz = params[key]
elif ts == 0:
pass
else:
print "full socket test required for ts:%d tgs:%d kernel:%d N:%d " % (
ts, tgs_r, kernel, N)
continue
for th in list(range(1, 1 + th_max)):
if (tgs > 1 and th % tgs != 0):
continue
machine_conf[
'pinning_args'] = "-m -g %s -C S1:0-%d -s 0x03 --" % (
group, (th - 1))
outfile = (
'kernel%d_isdp%d_ts%d_mwdt%d_tgs%d_N%d_th%d_%s_%s.txt'
% (kernel, is_dp, ts, mwdt, tgs_r, N, th, group,
exp_name[-13:]))
nt = max(
int(
float(th) / float(th_max) * k_time_scale[kernel] /
(N**3 / 1e6)), 30)
c_mwdt = mwdt if mwdt != -1 else 1
print outfile, th, tb, nwf, tgs, thx, thy, thz
run_test(dry_run=dry_run,
is_dp=is_dp,
th=th,
tgs=tgs,
thx=thx,
thy=thy,
thz=thz,
kernel=kernel,
ts=ts,
nx=N,
ny=N,
nz=N,
nt=nt,
outfile=outfile,
target_dir=target_dir,
cs=cs,
mwdt=c_mwdt,
tb=tb,
nwf=nwf)
count = count + 1
return count
def igs_test(target_dir, exp_name, th, group='', dry_run=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
cs = 8192
th = th
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if (machine_info['hostname'] == 'Haswell_18core'):
k_perf_order = {0: 150, 1: 500, 4: 40, 5: 200, 6: 20}
elif (machine_info['hostname'] == 'IVB_10core'):
k_perf_order = {0: 120, 1: 300, 4: 35, 5: 150, 6: 20}
k_time_scale = {
n: desired_time * k_perf_order[n]
for n in k_perf_order.keys()
}
#exp = is_dp, ts, k, N, bs_z, tb_l
exp_l = []
# spatial blocking
exp_l = exp_l + \
[(0, 0, 0, 960, 0, [-1])
,(1, 0, 0, 960, 0, [-1])
,(1, 0, 1, 960, 0, [-1])
,(1, 0, 4, 480, 0, [-1])
,(1, 0, 5, 680, 0, [-1])
]
# 1WD
exp_l = exp_l + \
[(0, 2, 0, 960, 1, [1, 3, 5])
,(1, 2, 0, 960, 1, [1, 3, 5])
,(1, 2, 1, 960, 1, [1, 3, 5, 7, 9, 11, 15, 19, 23, 29])
,(1, 2, 4, 480, 1, [1, 3, 5])
,(1, 2, 5, 680, 1, [1, 3, 9, 19])
]
# Solar kernel
exp_l = exp_l + \
[(1, 2, 6, 480, 1, [1, 3, 5, 7])
,(1, 2, 6, 480, 2, [1, 3, 5, 7])
,(1, 2, 6, 480, 3, [1, 3, 5, 7])
,(1, 2, 6, 480, 6, [1, 3, 5, 7])
,(1, 2, 6, 480, 9, [1, 3, 5, 7])]
mwdt = 1
tgs, thx, thy, thz = (1, 1, 1, 1)
count = 0
for is_dp, ts, kernel, N, bs_z, tb_l in exp_l:
for tb in tb_l:
outfile = ('kernel%d_isdp%d_ts%d_bsz$d_tb%d_N%d_%s_%s.txt' %
(kernel, is_dp, ts, bs_z, tb, N, group, exp_name[-13:]))
nt = max(int(k_time_scale[kernel] / (N**3 / 1e6)), 30)
# print outfile, ts, kernel, tb, N
run_test(ntests=1,
dry_run=dry_run,
is_dp=is_dp,
th=th,
tgs=tgs,
thx=thx,
thy=thy,
thz=thz,
kernel=kernel,
ts=ts,
nx=N,
ny=N,
nz=N,
nt=nt,
outfile=outfile,
target_dir=target_dir,
cs=cs,
mwdt=mwdt,
tb=tb,
nwf=bs_z)
count = count + 1
return count
def thread_scaling_test(target_dir, exp_name, group='', params={}, dry_run=0, is_tgs_only=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
is_dp=1
cs = 4096
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if(machine_info['hostname']=='Haswell_18core'):
k_perf_order = {0:1500, 1:5000, 4:400, 5:2000 ,6:100}
elif(machine_info['hostname']=='IVB_10core'):
k_perf_order = {0:1200, 1:3000, 4:350, 5:1500 ,6:80}
k_time_scale = {n: desired_time*k_perf_order[n] for n in k_perf_order.keys()}
points = dict()
points[0] = [896]
points[1] = [896]
points[4] = [768]
points[5] = [768]
points[6] = [384]
count=0
th_l = [machine_info['n_cores']]
th_max = machine_info['n_cores']
for ts, tgs_r in [(0,0), (2,1), (2,-1)]:
for kernel, mwdt in [(0,1), (1,2), (4,1), (5,2)]: #, (6,1)]:
for N in points[kernel]:
tb, nwf, tgs, thx, thy, thz = (-1,-1,tgs_r,-1,-1,-1)
# key = (mwdt, kernel, N, tgs_r, group)
# if key in params.keys():
# continue #already computed
if tgs==1: mwdt=-1
key = (mwdt, kernel, N, tgs_r, 'MEM')
if key in params.keys(): #reuse the conf of existing test
tb, nwf, tgs, thx, thy, thz = params[key]
elif ts==0:
pass
else:
print "full socket test required for ts:%d tgs:%d kernel:%d N:%d "%(ts, tgs_r, kernel, N)
continue
for th in list(range(1, 1+th_max)):
if (tgs>1 and th%tgs!=0):
continue
machine_conf['pinning_args'] = "-m -g %s -C S1:0-%d -s 0x03 --"%(group, (th-1))
outfile=('kernel%d_isdp%d_ts%d_mwdt%d_tgs%d_N%d_th%d_%s_%s.txt' % (kernel, is_dp, ts, mwdt, tgs_r, N, th, group, exp_name[-13:]))
nt = max(int( float(th)/float(th_max)* k_time_scale[kernel]/(N**3/1e6)), 30)
c_mwdt = mwdt if mwdt!=-1 else 1
print outfile, th, tb, nwf, tgs, thx, thy, thz
run_test(dry_run=dry_run, is_dp=is_dp, th=th, tgs=tgs, thx=thx, thy=thy, thz=thz, kernel=kernel, ts=ts, nx=N, ny=N, nz=N, nt=nt, outfile=outfile, target_dir=target_dir, cs=cs, mwdt=c_mwdt, tb=tb, nwf=nwf)
count = count+1
return count
def igs_test(target_dir, exp_name, th, group="", dry_run=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
cs = 8192
th = th
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if machine_info["hostname"] == "Haswell_18core":
k_perf_order = {0: 150, 1: 500, 4: 40, 5: 200, 6: 20}
elif machine_info["hostname"] == "IVB_10core":
k_perf_order = {0: 120, 1: 300, 4: 35, 5: 150, 6: 20}
k_time_scale = {n: desired_time * k_perf_order[n] for n in k_perf_order.keys()}
# exp = is_dp, ts, k, N, bs_z, tb_l
exp_l = []
# spatial blocking
exp_l = exp_l + [
(0, 0, 0, 960, 0, [-1]),
(1, 0, 0, 960, 0, [-1]),
(1, 0, 1, 960, 0, [-1]),
(1, 0, 4, 480, 0, [-1]),
(1, 0, 5, 680, 0, [-1]),
]
# 1WD
exp_l = exp_l + [
(0, 2, 0, 960, 1, [1, 3, 5]),
(1, 2, 0, 960, 1, [1, 3, 5]),
(1, 2, 1, 960, 1, [1, 3, 5, 7, 9, 11, 15, 19, 23, 29]),
(1, 2, 4, 480, 1, [1, 3, 5]),
(1, 2, 5, 680, 1, [1, 3, 9, 19]),
]
# Solar kernel
exp_l = exp_l + [
(1, 2, 6, 480, 1, [1, 3, 5, 7]),
(1, 2, 6, 480, 2, [1, 3, 5, 7]),
(1, 2, 6, 480, 3, [1, 3, 5, 7]),
(1, 2, 6, 480, 6, [1, 3, 5, 7]),
(1, 2, 6, 480, 9, [1, 3, 5, 7]),
]
mwdt = 1
tgs, thx, thy, thz = (1, 1, 1, 1)
count = 0
for is_dp, ts, kernel, N, bs_z, tb_l in exp_l:
for tb in tb_l:
outfile = "kernel%d_isdp%d_ts%d_bsz$d_tb%d_N%d_%s_%s.txt" % (
kernel,
is_dp,
ts,
bs_z,
tb,
N,
group,
exp_name[-13:],
)
nt = max(int(k_time_scale[kernel] / (N ** 3 / 1e6)), 30)
# print outfile, ts, kernel, tb, N
run_test(
ntests=1,
dry_run=dry_run,
is_dp=is_dp,
th=th,
tgs=tgs,
thx=thx,
thy=thy,
thz=thz,
kernel=kernel,
ts=ts,
nx=N,
ny=N,
nz=N,
nt=nt,
outfile=outfile,
target_dir=target_dir,
cs=cs,
mwdt=mwdt,
tb=tb,
nwf=bs_z,
)
count = count + 1
return count
def igs_test(target_dir, exp_name, th, group='', params={}, dry_run=0, is_tgs_only=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
is_dp=1
cs = 4096
th = machine_info['n_cores']
k_r = {0:4, 1:1, 4:4, 5:1, 6:1}
kernels_min_limits = {n: k_r[n]*th*4 for n in k_r.keys()}
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if(machine_info['hostname']=='Haswell_18core'):
tgs_l = [2,3,6,9,18]
increment = 128
k_perf_order = {0:1500, 1:5000, 4:400, 5:2000 ,6:100}
if is_dp == 1:
kernels_limits = {0:1281, 1:1409, 4:769, 5:897, 6:513}
else:
kernels_limits = {0:2100, 4:1200}
elif(machine_info['hostname']=='IVB_10core'):
tgs_l = [2,5,10]
increment = 64
k_perf_order = {0:1200, 1:3000, 4:350, 5:1500 ,6:80}
if is_dp ==1:
kernels_limits = {0:1089, 1:1217, 4:577, 5:705, 6:289}
else:
kernels_limits = {0:1350, 4:801}
points = dict()
points[0] = [64] + list(range(128, 5000, increment))
points[1] = points[0]
points[4] = points[0]
points[5] = points[0]
points[6] = points[0]
k_time_scale = {n: desired_time*k_perf_order[n] for n in k_perf_order.keys()}
if(is_tgs_only):
exp_l = [(2,tgs_l)]
else:
exp_l = [(2,[-1, 1]), (0,[0])]
count=0
for ts, tgs_rl in exp_l:
for tgs_r in tgs_rl:
for kernel, mwdt_list in [(0,[1]), (1,[2]), (4,[1]), (5,[2])]:
# for kernel, mwdt_list in [(6,[1])]:
if ts==0 or tgs_r==1:
mwdt_list=[-1]
for mwdt in mwdt_list:
for N in points[kernel]:
if( ((tgs_r!=1) or (N >= kernels_min_limits[kernel])) and (N < kernels_limits[kernel]) ):
tb, nwf, tgs, thx, thy, thz = (-1,-1,tgs_r,-1,-1,-1)
key = (mwdt, kernel, N, tgs_r, group)
if key in params.keys():
print "Already computed:", key
continue #already computed
key = (mwdt, kernel, N, tgs_r, 'MEM')
if key in params.keys(): #reuse the conf of existing test
tb, nwf, tgs, thx, thy, thz = params[key]
outfile=('kernel%d_isdp%d_ts%d_mwdt%d_tgs%d_N%d_%s_%s.txt' % (kernel, is_dp, ts, mwdt, tgs_r, N, group, exp_name[-13:]))
nt = max(int(k_time_scale[kernel]/(N**3/1e6)), 30)
c_mwdt = mwdt # to avoid negative array access at the c code
if mwdt==-1: c_mwdt=1
# print outfile, tb, nwf, tgs, thx, thy, thz
run_test(dry_run=dry_run, is_dp=is_dp, th=th, tgs=tgs, thx=thx, thy=thy, thz=thz, kernel=kernel, ts=ts, nx=N, ny=N, nz=N, nt=nt, outfile=outfile, target_dir=target_dir, cs=cs, mwdt=c_mwdt, tb=tb, nwf=nwf)
count = count+1
return count
def igs_test(target_dir,
exp_name,
th,
group='',
params={},
dry_run=0,
is_tgs_only=0):
from scripts.conf.conf import machine_conf, machine_info
from scripts.utils import run_test
import itertools
is_dp = 1
cs = 4096
th = machine_info['n_cores']
k_r = {0: 4, 1: 1, 4: 4, 5: 1, 6: 1}
kernels_min_limits = {n: k_r[n] * th * 4 for n in k_r.keys()}
# Test using rasonable time
# T = scale * size / perf
# scale = T*perf/size
desired_time = 20
if (machine_info['hostname'] == 'Haswell_18core'):
tgs_l = [2, 3, 6, 9, 18]
increment = 128
k_perf_order = {0: 1500, 1: 5000, 4: 400, 5: 2000, 6: 100}
if is_dp == 1:
kernels_limits = {0: 1281, 1: 1409, 4: 769, 5: 897, 6: 513}
else:
kernels_limits = {0: 2100, 4: 1200}
elif (machine_info['hostname'] == 'IVB_10core'):
tgs_l = [2, 5, 10]
increment = 64
k_perf_order = {0: 1200, 1: 3000, 4: 350, 5: 1500, 6: 80}
if is_dp == 1:
kernels_limits = {0: 1089, 1: 1217, 4: 577, 5: 705, 6: 289}
else:
kernels_limits = {0: 1350, 4: 801}
points = dict()
points[0] = [64] + list(range(128, 5000, increment))
points[1] = points[0]
points[4] = points[0]
points[5] = points[0]
points[6] = points[0]
k_time_scale = {
n: desired_time * k_perf_order[n]
for n in k_perf_order.keys()
}
if (is_tgs_only):
exp_l = [(2, tgs_l)]
else:
exp_l = [(2, [-1, 1]), (0, [0])]
count = 0
for ts, tgs_rl in exp_l:
for tgs_r in tgs_rl:
for kernel, mwdt_list in [(0, [1]), (1, [2]), (4, [1]), (5, [2])]:
# for kernel, mwdt_list in [(6,[1])]:
if ts == 0 or tgs_r == 1:
mwdt_list = [-1]
for mwdt in mwdt_list:
for N in points[kernel]:
if (((tgs_r != 1) or (N >= kernels_min_limits[kernel]))
and (N < kernels_limits[kernel])):
tb, nwf, tgs, thx, thy, thz = (-1, -1, tgs_r, -1,
-1, -1)
key = (mwdt, kernel, N, tgs_r, group)
if key in params.keys():
print "Already computed:", key
continue #already computed
key = (mwdt, kernel, N, tgs_r, 'MEM')
if key in params.keys(
): #reuse the conf of existing test
tb, nwf, tgs, thx, thy, thz = params[key]
outfile = (
'kernel%d_isdp%d_ts%d_mwdt%d_tgs%d_N%d_%s_%s.txt'
% (kernel, is_dp, ts, mwdt, tgs_r, N, group,
exp_name[-13:]))
nt = max(int(k_time_scale[kernel] / (N**3 / 1e6)),
30)
c_mwdt = mwdt # to avoid negative array access at the c code
if mwdt == -1: c_mwdt = 1
# print outfile, tb, nwf, tgs, thx, thy, thz
run_test(dry_run=dry_run,
is_dp=is_dp,
th=th,
tgs=tgs,
thx=thx,
thy=thy,
thz=thz,
kernel=kernel,
ts=ts,
nx=N,
ny=N,
nz=N,
nt=nt,
outfile=outfile,
target_dir=target_dir,
cs=cs,
mwdt=c_mwdt,
tb=tb,
nwf=nwf)
count = count + 1
return count
